Abstract

Keywords

Introduction (Jung)

What do we lready know about diabetes

Research question?

Methods

Dataset (Jung)

descriptions, distributions, missing data

Dataset versions (Jung)

original, naomit, mice, skip

Measures

For the assessment of the model in our experiment we will use three commonly known measures:

  1. AUC - Area Under ROC Curve

An ROC curve (receiver operating characteristic curve) is a graph that shows the performance of a classification model at all classification thresholds. This curve plots two parameters: TPR (True Positive Rate) and FPR (False Positive Rate), defined as:

\(TPR=\frac{TP}{TP+FN}\)

\(FPR=\frac{FP}{FP+TN}\)

  1. BAC - BAlanced Accuracy

This measure is a weighted accuracy, that is more suitable for datasets with imbalanced classes.

  1. FNR - False Negative Rate

It corresponds to the number of cases where model incorrectly fails to indicate the presence of a condition when it is present, divided by the number of all.

\(FNR=\frac{FN}{TP+FN}\)

Models

To identify the models, that could potentially be efficient classificators on this dataset, we measured the AUC, BAlanced Accuracy and False Negative Rate measures on 6 different classification models from mlr package:

  • Ranger - Random regression forest

  • Ada - Adaptive Boosting

  • Binomial - Binomial Regression

  • GBM - Gradient Boosting Machine

  • Naive Bayes - Naive Bayes

  • Random Forest - Random Forest

Explanations (Karol)

In this section we will take time to investigate some model explanations that we will be using later. Feel free to use this chapter as reference for quick overview. For more in-depth knowledge please refer to https://ema.drwhy.ai/InstanceLevelExploration.html. //bookdown?? TODO All images are also taken from the article.

Local

Local explanations allow us to better understand model’s prediction from the level of single observations. This can be useful when we want to evaluate predictions for particular instances. For example, which values have the most importance for the particular observations, or how the change in variables would impact the result. Local explanations combined with professional expertise could sometimes hint towards potential problems with the model in case they contradict each other.

Break Down

In this explanation all observations start from the same base that is the mean prediction for all data. Next steps consist of fixing consecutive variables on values from the observation in question and measuring the change in mean prediction that is now calculated with some of the values fixed. This shift in mean prediction is interpreted as the impact of this variable’s value on prediction for this observation. Figure TODO shows the corresponding steps.

Figure: Panel A shows the distribution and mean prediction with values on the same level and above fixed on values from observation. Then change in mean is calculated and assigned to corresponding variables.

Figure: Panel A shows the distribution and mean prediction with values on the same level and above fixed on values from observation. Then change in mean is calculated and assigned to corresponding variables.

A big drawback for this method is how the order of variable can influence the explanation outcome. This problem occurs most often as a result of existing interactions between variables as seen in OBRAZEK TODO

Figure: Variables age and class interact with each other so the changing order results in different plots

Figure: Variables age and class interact with each other so the changing order results in different plots

Shap

Shap is a direct response to the biggest problem of Break Down method that is the ordering chosen for explanation could alter the results as seen in figure TODO. Shap takes a number of permutation and calculates the average response of Break Down on these permutations. More permutations result in more stable explanations. Exemplary output is shown in figure TODO.

Figure: Different results of Break Down for 10 random variable orders

Figure: Different results of Break Down for 10 random variable orders

Figure: Shap output. Average impact calculated with Break Down for 10 permutations

Figure: Shap output. Average impact calculated with Break Down for 10 permutations

Lime

The essence of Lime is to locally approximate the black box model with a glass box one. Then we can use it for local explanations that should yield results appropriate for the black box. Firstly, we generate data “close” to our observation (also called Point of Interest) and make predictions with black box model. Based on those predictions we try to fit the glass box model that accurately predicts those observations. In result we receive a glass box model that should behave the same as original model as long as we remain close to the Point of interest. Figure TODO describes the idea.

Figure: Colored areas correspond to different prediction for the black box model. Black cross is our Point of Interest and circles are generated data. The line represents our simpler model that approximates original model around Point of Interest

Figure: Colored areas correspond to different prediction for the black box model. Black cross is our Point of Interest and circles are generated data. The line represents our simpler model that approximates original model around Point of Interest

This method is often used for datasets consisting of many variables. Methods like Break down and Shap come out short in these situations.

Ceteris Paribus

Ceteris Paribus is a Latin phrase meaning “all else unchanged”. This accurately describes what it does. For the certain observation we take values that are interesting to us and observe how the prediction changes as we change values in those columns one at a time. This can yield interesting conclusions about change in values and its impact on predictions. Figures TODO and TODO show prediction change as responce for variable change.

Figure: Shape of lines resebles the nature of the model. Dot shows value and model’s prediction for the observation.

Figure: Shape of lines resebles the nature of the model. Dot shows value and model’s prediction for the observation.

Figure: Ceteris Paribus shows different importance in variables and their changes for different observations.

Figure: Ceteris Paribus shows different importance in variables and their changes for different observations.

Ceteris Paribus is very popular due to its simplicity and interpretability. The biggest drawback however is the possibility of unpredictable or misleading results. For example, with fixed variable age set to 18, predicting outcome for a significant number of pregnancies does not make sense. It is also hard to capture interactions when working with variables separately.

Global

As the name implies in global explanations we look at variables from the perspective of whole dataset. We can check the importance of variable for the model or analyze average impact of certain values on predictions.

Feature Importance

The main idea behind Feature Importance is measuring how important each variable is for model’s predictions. To do that we measure the change in AUC after permutating values of each variable. The bigger the change, the bigger importance of variable. To stabilize the results, we can measure average change for a number of permutations. In figure TODO we can see that variable gender is the most important for model’s prediction.

Figure: Feature Importance output for 10 permutations.

Figure: Feature Importance output for 10 permutations.

Partial Dependence

To put it short, Partial Dependence is the average of Ceteris Paribus for whole data. This results in average importance and impact of variable and its values. The similarities can be observed in figure TODO

Figure: Multiple observations with Ceteris Paribus and their average as Partial Dependence.

Figure: Multiple observations with Ceteris Paribus and their average as Partial Dependence.

As we can see observations can return different shapes for Ceteris Paribus, hence averaging them could mean loss of information. To answer this problem Partial Dependence implements grouping and clustering that could show the difference as shown in figure TODO.

Figure: Partial Dependence grouped by sex.

Figure: Partial Dependence grouped by sex.

As in Ceteris Paribus this method is very simple and understandable. However, it also carries over the same problems resulting from correlated variables.

Accumulated Dependence

Also referred as Accumulated-Local Dependence. It is the direct answer for correlation issue in Partial Dependence. The construction of Accumulated Dependence is the same as in Ceteris Paribus. The only difference being how the observations are summarized. Partial Dependence uses marginal distribution while Accumulated Dependence uses conditional distribution. This means that when there are at most negligible correlations, Accumulated Dependence yields results very similar to Partial Dependence.

Results

Model comparison

The models were checked on both original and skip dataset version, with the following results:

Data Original

The measure values between models are very similar, however,

Data Skip

Based on the results, two models: Ranger and Ada where chosen for further analysis. Hyperparamether tuning using random grid search was peformed, resulting in classificators with the following values:

Ranger

  • num.trees=776,
  • mtry=1,
  • min.node.size=8
  • splitrule=“extratrees”

Ada - loss=‘logistic’ - type=‘discrete’, - iter=81, - max.iter=3, - minsplit=45, - minbucket=4, - maxdepth=1

Explanations results

Global

Variable Importance

Figure TODO shows Feature Importance for booth Ada and Ranger models. As we can see, booth aggree with eachother as to what variables are most important to them. Their predictions are mostly based on the values of plas variable and then mass and age come second.

Figure: Variable importance for Ada and Ranger based on AUC loss after permutations.

Figure: Variable importance for Ada and Ranger based on AUC loss after permutations.

Partial Dependence Profile

Partial Dependence profile seen in figure TODO give us perhaps the most interesting results. Firstly, they are adequate to Feature importance in Figure TODO. The more important variables return wider range of average prediction and less important give less varied predictions respectively. Secondly, we can observe the difference in model’s inner characteristics based on the shape of their’s PDP functions. Ada is more edgy and straight while ranger more smooth. Based on results yielded by PDP and acording to these two models we can conclude that, generally speaking the bigger the values the higher the chance of positive Diabetes test result.

Some other key observations include:

  • Little to none impact by change of age after 35 years of living.
  • A sudden increase in prediction with mass reaching 30 being the border of obesity in BMI model.
  • Almost no change for ada prediction with change of pedi.
  • A big increase in prediction with 7th pregnancy.
Figure: Average predictions with PDP for ada and ranger.

Figure: Average predictions with PDP for ada and ranger.

Figure: Partial dependence and acumulated dependence on ranger’s mass variable.

Figure: Partial dependence and acumulated dependence on ranger’s mass variable.

Local

Discussion

Expert opinion

There are two main types of diabetes:

  • type 1 diabetes, where the body does not make insulin. It is mostly caused by genetic and enviromental factors

  • type 2 diabetes, where the body does not make or use insulin well. It is the most common kind of diabetes. It is caused by several factors, including lifestyle factors and genes

Unfortunately, we do not have any data about the type of diabetes that was found among the patients in the investigated dataset. However, since type 2 diabetes is the most common type, and some of the factors causing the different types are common, we will focus mostly on the risk factors for type 2 diabetes.

Causes of type 2 diabetes

According to the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK), type 2 diabetes can be caused by several factors:

  • obesity and physical inactivity

Extra weight sometimes causes insulin resistance and is common in people with type 2 diabetes.

  • insulin resistance

Type 2 diabetes usually begins with insulin resistance, a condition in which muscle, liver, and fat cells do not use insulin well. As a result, your body needs more insulin to help glucose enter cells. At first, the pancreas makes more insulin to keep up with the added demand. Over time, the pancreas can’t make enough insulin, and blood glucose levels rise.

  • genes and family history

As in type 1 diabetes, certain genes may make you more likely to develop type 2 diabetes. Genes also can increase the risk of type 2 diabetes by increasing a person’s tendency to become overweight or obese.

  • ethnicity

Diabetes occurs more often in these racial/ethnic groups: - African Americans - Alaska Natives - American Indians - Asian Americans - Hispanics/Latinos - Native Hawaiians - Pacific Islanders

  • age Type 2 diabetes occurs most often in middle-aged and older adults, but it can also affect children.

  • gestational diabetes Gestational diabetes is the type of diabets thatdevelops during pregnancy and is caused by the hormonal changes of pregnancy along with genetic and lifestyle factors. Women with a history of gestational diabetes have a greater chance of developing type 2 diabetes later in life.

Comparison

The causes of diabetes according to experts directly correspond with the columns is the explained dataset:

  • plas - plasma glucose concentration, can indicate whether tha patient suffers from a insulin resistance
  • mass - Body Mass Index (BMI) can be one of the indicators whether patients weight puts him at risk for type 2 diabetes
  • age- older people are generally more likely to develop diabetes
  • pedi - predigree function is meant to be an indicator of patients risk for diabetes based on his genes and family history
  • preg - women that have been pregnant multiple times are more likely to develop gestational diabetes, and therefor, type 2 diabetes later in life.

Since all the women in the dataset have the same ethnicity, it is not a factor that can be considered valuable in this experiment.

Conclusion

References